Improve speed and accuracy

Update VMMR (Vehicle Make Model Recognition) and LPCI (License Plate …

…Country Identification ) to improve accuracy

Fix issue #241, #237, #238, #234, #171, #159, #144

This version consumes 3+ times less CPU and GPU memory than the previous version (3.7.x)
We improved the accuracy for VMMR and LPCI models.
More car make/model added

This version mainly adds support for "chinese" license plates and imp…

…roves OCR accuracy for "latin" and "korean" charsets.

The online sample for China is at https://www.doubango.org/webapps/alpr-china/

Fix issue #227, #226, #213, #203, #202, #200, #198, #196, #195, #194, #192, #188, #180, #170, #169, #145, #151, #154, #131

- Provide commercial license for Embedded Linux AArch64
- License generation: Use configured CPUs instead of online/available ones
- Improve accuracy in Brazilian plates
- Improve accuracy for Taiwanese license plate
- Restrict number of jobs (JSON config "max_jobs") to avoid CompV drowning due to the number of threads
- CompV drowning due to the number of threads (#144)
- Add C++ function to the API to retrieve EXIF orientation value
- Should not return error when loading OpenVINO plugin fails
- Disable OpenVINO on runtimeKey generation
- Improve detection accuracy on orphan plates (dark night)
- Make IENV disabled by default
- Make memory pooling more versatile (release forks faster)
- Autolock the process function to enforce thread safety
- Use static linking for libcurl for AWS/Azure activations
- Add more Indonesian motorcycle plates in the detection and recognition training set
- Add '육' in training set for Korean plates
- Make rectification enabled by default (Java, Python, C++, C#) and all platform (Android, RPI...)

Mainly updating Tensorflow lite version to v2.5.0-rc1. Only Android a…

…nd Raspberry Pi are impacted.

Fix issue #143, #140, #139 and #135

Fix issue #132, #127, #126, #124, #123, #122, #118 and #61

+ We have improved the accuracy for the detector.

    - Now we detect motorcycles in addition to cars

    - You'll have fewer false positives on cars without plates.

    - The bounding boxes on overlapping cars are more accurate.

+ We have improved the OCR models.

+ We have improved the accuracy for the Pyramidal search. The bounding boxes on far away or small cars and plates are tighter.

+ We have improved the accuracy for Vehicle Body Style Recognition (VBSR) and added 'motorcycle' to the classes.

+ We have improved the accuracy for Vehicle Color Recognition (VCR).

+ Completely redesigned License Plate Country Identification (LPCI) to improve the accuracy.

+ Return "frame_id" in every callback when parallel mode is enabled to ease matching the input to the result.

+ Now it's possible to attach the license to a Microsoft Azure VM instead of hardware to avoid requiring a dedicated server. More info at https://github.com/DoubangoTelecom/ultimateALPR-SDK/blob/master/AWS.md#azure

+ Now we return the OCR score for each character/digit in the license plate number. More info about the JSON 'confidences' field at https://www.doubango.org/SDKs/anpr/docs/Frequently_Asked_Questions_(FAQ).html#could-you-explain-what-each-field-in-the-json-result-mean. This is backward compatible, no change needed in your code.

*** Version 3.2.0 ***

- Fix issue #117, #115, #114, #99, #97 #92

1/ Added support for Image Enhancement for Night-Vision (IENV) using AI. This feature improves the image quality when you're filming at night. No special harward is require, it works with any camera. Here is a video showing the output of the filter: https://www.youtube.com/embed/QCkLPP1ix-c
More info: https://www.doubango.org/SDKs/anpr/docs/Features.html#image-enhancement-for-night-vision-ienv

2/ Added support for Vehicle Body Style Recognition (VBSR). On NVIDIA Jetson you'll need to generate the plan for this new model. The process will be fast as existing models will be ignored if the corresponding plan exist.
More info: https://www.doubango.org/SDKs/anpr/docs/Features.html#vehicle-body-style-recognition-vbsr

3/ Added more training samples to improved LPR/OCR accuracy

4/ It's now possible to ask the SDK to process and return cars with no plate. JSON configuration entry: https://www.doubango.org/SDKs/anpr/docs/Configuration_options.html#car-noplate-detect-enabled

5/ On ARM, the rectification layer is now fully written in assembler with NEON SIMD acceleration. We recommend enabling this feature on ARM to improve the accuracy for skewed or slanted plates.
More info: https://www.doubango.org/SDKs/anpr/docs/Rectification_layer.html